EAGLE3 offline disaggregated parity

This page preserves a historical correctness check from commit 40d8feff. It is evidence that the offline producer/consumer path and colocated path were numerically aligned in that experiment, not a current-release performance or quality claim.

Qwen2.5-7B, two H200 nodes

The comparison used the same Qwen2.5-7B model build, precomputed feature tensors, and random seed for both paths. The disaggregated consumer and the colocated baseline assembled the EAGLE3 runtime through the same training implementation. Feature tensors were byte-identical; the remaining metric differences were approximately 1e-6 to 1e-8 GPU numerical noise.

Step	Metric	Disaggregated	Colocated
20	acceptance_rate	0.0013300	0.0013300
20	ploss	5.386736	5.386740
20	acc	0.0272590	0.0272590
120	acceptance_rate	0.0223610	0.0223505
180	acceptance_rate	0.0337013	0.0336982

Both accuracy and the training-time acceptance proxy increased during the run. Per-step values remained noisy because the experiment used batch size 1 over 64 diverse samples. This proxy is useful for path parity; it is not the same measurement as serving-time accepted length.

Unified entry for a new run

The current equivalent recipe is examples/configs/qwen2.5-7b-eagle3-offline-disaggregated.yaml. To place the offline producer and consumer on different nodes, run one checked-in command through the cluster launcher:

rcli exec --per-node <job> \
  'CONFIG=examples/configs/qwen2.5-7b-eagle3-offline-disaggregated.yaml bash examples/disagg/run_offline_2node.sh'

Rank 0 invokes specforge train --role producer; rank 1 invokes specforge train --role consumer. Both nodes must resolve the config’s control_dir, store_root, hidden-state input, and vocabulary mapping to the same data. Use a fresh attempt directory, then compare against the colocated examples/configs/qwen2.5-7b-eagle3-offline.yaml recipe with the same inputs, seed, and training overrides.